Application of Data-driven Approaches to Efficient Knock Prediction in Engine Development

Designing environmentally sound, efficient, and robust large internal combustion engines involves balancing a large number of desirable and undesirable combustion properties. Knocking—a phenomenon of uncontrolled and abnormal combustion—is one such undesirable characteristic of sparkignited engines. During the design process, it is therefore necessary to determine the knock tendency in relation to the corresponding engine operating point. To this end, single-cylinder engine tests are important as they allow the evaluation of a variety of engine operating points for knocking under controlled conditions. But even with single-cylinder engine testing, experimental resources are limited. For this reason, a well-structured approach is required in order to efficiently determine the knock tendency.
This study presents a method for predicting the engine knock probability of an operating point based on a limited amount of measurement data. A regression model is built that describes the (empirical) knock probability as a function of knocking-relevant control variables and additional relevant engine operating parameters. The required training data is generated from engine measurements that also determine the parameter space limits of knocking-relevant control variables such as ignition timing or charge air pressure. Eventually, this model allows prediction of the knock probability of operating points that are not studied during the engine tests. The presented method is verified by multiple engine tests on a large single-cylinder engine fueled with natural gas or hydrogen. The high prediction accuracy of the empirical knock probability for unseen points demonstrates the potential benefit of the approach. Since it is not tied to a specific engine size or type, this method can be used in the design process of other engines as well.